| Strawberry is a perennial herb which is heart-shaped in appearance,high in sugar and sweet in taste.Strawberries have high nutritional value and are rich in carotene,glucose,malic acid,dietary fiber,various vitamins such as A,B1,B2,C,E,and various minerals such as calcium,potassium,iron,sodium,and magnesium.As high economic value crops,strawberries have been grown in large numbers around the world.However,strawberry growers are always lacking some efficient and intelligent meathods to quickly detect strawberry growth and predict strawberry yield.This paper combined UAV(Unmanned Aerial Vehicle)remote sensing multispectral technology,deep learning networks,object detection algorithms,3D reconstruction and oblique photography technology to develop a series of UAV remote sensing platforms and devices to help growers detect strawberry growth intelligently.The main conclusions were as follows:(1)An environment-controllable five DOF(degree-of-freedom)UAV low-altitude remote sensing simulation platform was developed.The platform could simulate horizontal,vertical,pitch,tilt,and roll motions of the UAV.Using this simulation platform equipped with a multispectral camera to obtain strawberry seedling canopy multispectral images at different simulated flight heights and speeds.Extracting 6 vegetation indexes to establish SPAD linear prediction models and explore the impact of image acquisition speed and height on the predictive performance of the models.The best linear prediction model of vegetation index was finally selected to establish a visual inversion map of strawberry canopy under different nitrogen stress.The results show that when the camera height was 2.0 m and the simulated flight speed was 0.1m/s,the linear prediction model based on the vegetation index NIR/G had the best effect with correlation coefficient Rp equal to 0.7481.The SPAD value of the excess nitrogen group and the appropriate nitrogen group were significantly higher than the deficiency nitrogen group which was consistent with the actual measurement results.(2)An automatic real-time detection system for flowers and fruits in strawberry fields was developed based on the UAV remote sensing platform and object detection algorithm.In order to compare the impact of different flight heights on the detection results,the system used two different heights of 2 m and 3 m to acquire images.The detection results showed that Sensation varieties of ripe strawberries captured at 2m height had the best detection effect with average precision(AP)equal to 0.91,and the Radiance variety of immature strawberries captured at 3m height was the most difficult to detect with an average precision of 0.61.Using this system to count the number of flowers in the field on different experimental dates.The results were compared with the artificial statistical numbers and it showed that the average accuracy of the flower count of the system was 84.1%,and the average occlusion rate of flowers in the field was13.5%.Finally,the system was used to establish the yield distribution map of strawberry flowers and fruits on different dates which could help growers predict the incoming yield of the strawberry field.(3)Proposed Straw R-CNN object detection algorithm which was based on the Faster RCNN framework.The Straw R-CNN modified the backbone network for extracting feature maps and added a feature pyramid structure and a Ro I Aline layer to improve the detection accuracy of small-sized targets such as strawberry flowers.The Straw R-CNN object detection algorithm was compared with other region-based object detection algorithms including Faster R-CNN,Fast RCNN,and R-CNN algorithms by training and detecting the same strawberry remote sensing dataset.The results showed that Straw R-CNN object detection algorithm was significantly better than the other three algorithms in terms of detection accuracy(0.772 m AP),training time(5.5hours),and detection speed(8.850 FPS).(4)Developed the ZTRS-M5 B small UAV oblique photography system.Using this system to collect images of large-scale strawberry fields and establish orthoimages,the results showed that,compared with the single len UAV,the system can establish orthoimages of large-scale strawberry fields more quickly with higher accuracy.Two complex regional environments in a rural residential area and a factory area were selected to further test the performance of this oblique photography system.The accuracy of the model was verified by comparing the measured 3D model values with ground truth values.The results met remote sensing mapping accuracy requirements.Compared to other oblique photography systems on the market,the ZTRS-M5B has more advantages in cost and portability. |
